fix(flow-chat): report token throughput from streaming time

[bobleer]

Summary

• The per-turn t/s readout divided turn-level output tokens by turn wall time, which also covers tool execution, permission waits and retries. On any tool-heavy turn the model looked an order of magnitude slower than it actually generates.
• Throughput now divides by the streaming window. getTurnOutputDurationMs sums, per model round, streamDurationMs - firstChunkMs (falling back to firstVisibleOutputMs), so request handshake, provider queueing and prompt prefill are not charged to generation. Summing across the turn's rounds keeps the denominator aligned with the turn-level output token count that model rounds already accumulate.
• Both fields already came from the backend (stream_duration_ms on ModelRoundCompleted) and are persisted/rehydrated; the frontend simply never used them for the rate. Rounds without stream timing (older persisted sessions) keep the previous wall-time fallback instead of reporting an invented number.
• The turn elapsed time in the completion metadata is unchanged — a duration is not throughput.
• modelRound.metrics.rateBasis is updated in en-US, zh-CN and zh-TW so the tooltip no longer claims the rate includes tool time.

Fixes the inaccurate X t/s value shown in each turn footer.

Type and Areas

Type: bug fix (UI metric accuracy)

Areas: web UI (flow_chat projection, model round footer, locales)

Motivation / Impact

A turn that spent 90s in tools and 10s streaming 2,000 output tokens displayed roughly 22 t/s instead of 200 t/s. The number is a user-facing quality signal for model/provider speed, so it needs to measure generation, not turn duration.

No persisted shape changes: no new fields on disk, no migration, no config surface. Older sessions without stream_duration_ms behave exactly as before.

Verification

• pnpm --dir src/web-ui run test:run src/flow_chat/utils/turnUsageMetrics.test.ts src/flow_chat/store/modernFlowChatStore.test.ts — passed (114 tests)
• pnpm --dir src/web-ui run test:run src/flow_chat/components/modern src/flow_chat/utils/tokenUsageDisplay.test.ts — passed (699 tests)
• pnpm run type-check:web — passed
• pnpm run i18n:audit — passed (0 warnings)
• eslint on the eight changed source files — passed (no errors)

New cases cover: first-chunk subtraction, visible-output fallback, whole-window fallback when no first-output mark exists, multi-round summation, inverted timing reporting nothing instead of a negative window, and missing stream timing yielding undefined so the caller falls back.

Reviewer Notes

• check:core-boundaries / remote-surface registry are untouched: this is a display-only projection change with no new Tauri command or persisted field.
• Remote scenarios: no change to protocol, persisted data, or peer/dispatch surfaces. Remote control surfaces render the same footer through the same projection, so they inherit the corrected rate without extra work.
• Observable change for reviewers: with the corrected denominator the same turn reads much higher (e.g. 8 t/s → 200 t/s), so the speedLevel tier also moves up. The 15/40/80 boundaries were calibrated against the old, deflated numbers, so a normal provider now lands in fast/veryFast more often. Worth a second opinion on whether those thresholds should be retuned in a follow-up; this PR keeps them unchanged so the only variable is the denominator.
• The t/s value is still absent while a round is actively streaming (streamDurationMs is only known at round completion). That is pre-existing behaviour, not a regression of this PR.

Checklist

• This PR is focused and does not include secrets, temporary prompts, generated scratch files, or unrelated artifacts.
• Relevant verification is recorded above, or skipped checks are explained.
• User-facing strings, docs, and locales are updated where applicable.

Generated with OpenBitFun